Opus 4.7 breaks only two prompt patterns, not all of them
Phil Rentier Digital argues that Opus 4.7 didn't break most prompts — it simply stopped silently patching two specific gaps that 4.6 used to fill automatically: implicit tool calls and implicit scope exceptions.
Score breakdown
Developers maintaining `CLAUDE.md` files or system prompts for Claude-based agents can avoid unnecessary rewrites by targeting only two specific patterns — non-binding action verbs on tool-dependent steps and scope rules without explicit exceptions — rather than auditing every prompt from scratch.
- 01Anthropic's prompting best practices page states that prompts written for 4.6 run very well on 4.7 out of the box — a line the post says most migration guides omitted.
- 02Opus 4.7 dropped on April 16th, triggering a wave of migration guides the post characterizes as recycling the same four bullet points from Anthropic's docs.
- 034.6 silently inferred tool calls from soft verbs like 'check', 'consider', 'look at', and 'review'; 4.7 treats these as mental suggestions, not tool invocations.
Phil Rentier Digital's post pushes back against the wave of Opus 4.7 migration guides that appeared within hours of the model's April 16th release, arguing they all recycled the same four bullet points from Anthropic's documentation while missing a critical sentence: Anthropic explicitly states that prompts written for 4.6 run very well on 4.7 out of the box. The real change, the post argues, is a single behavioral posture shift — 4.7 reads instructions literally rather than cooperatively filling in gaps.
The post identifies exactly two gap types that 4.6 used to patch silently.
The post identifies exactly two gap types that 4.6 used to patch silently. First, implicit tool calls: verbs like "check", "consider", "look at", and "review" were treated by 4.6 as instructions to actually invoke tools (e.g., a `Glob` or `Read` call), whereas 4.7 treats them as suggestions to reason mentally. A `CLAUDE.md` rule like "before suggesting code changes, check the current project structure" would reliably trigger a `Glob` call on 4.6 but produce no tool invocation on 4.7, causing the model to propose changes from memory. Second, implicit scope and exceptions: rules like "warn the customer about restocking delays once per conversation" or "never use exclamation marks" were interpreted by 4.6 with inferred common-sense boundaries — reopening the warning for a new order, or not stripping punctuation from verbatim quoted content. 4.7 applies such rules exactly as written, with no inferred exceptions.
The prescribed fix is narrow: audit `CLAUDE.md` and system prompts only for non-binding action verbs on load-bearing steps, and make tool calls explicit; separately, add explicit exception clauses to any scope-limiting rules. The post frames everything else in the migration discourse — faster reasoning, fewer default tool calls, more direct tone — as downstream consequences of this one posture change, not independent breaking changes requiring separate remediation.
Key facts
- 01Anthropic's prompting best practices page states that prompts written for 4.6 run very well on 4.7 out of the box — a line the post says most migration guides omitted.
- 02Opus 4.7 dropped on April 16th, triggering a wave of migration guides the post characterizes as recycling the same four bullet points from Anthropic's docs.
- 034.6 silently inferred tool calls from soft verbs like 'check', 'consider', 'look at', and 'review'; 4.7 treats these as mental suggestions, not tool invocations.
- 044.6 inferred common-sense exceptions to scope rules (e.g., 'once per conversation' reopening for a new order); 4.7 applies such rules exactly as written.